A language model maps context to a prediction or structured proposal. Tools, state, policies, validation, and control flow turn model calls into a bounded agent system.

A Model Is One Component, Not the Whole Agent

In this situation—asking a study helper to explain fractions without giving away the final answer—which choice best applies “Instructions Constrain Behavior”?

A prompt is an input contract for a model node. Treating it as versioned configuration makes its role, assumptions, output schema, and evaluation history visible.

Prompts Are Versioned Node Inputs

An instruction narrows the actions a system should take. Useful instructions identify the task, boundaries, relevant context, and the form of an acceptable result.

Goals, input and output, feedback, memory, choice, uncertainty, graphs, human control, privacy, learning, and failure—without assumed technical background.

Everyday Foundations

DeepLearning.AI

Canonical URL: https://www.deeplearning.ai/courses/agentic-ai
License: Publicly viewable; no open-content license asserted
Restrictions: Link and independently summarize ideas only; do not reproduce video or transcript.

Agentic AI with Andrew Ng — Graph Engineering Course Source

A decision compares available actions using current information and a criterion. In a graph, a decision node controls which outgoing edge is followed.

A Choice Selects Among Alternatives

An agent goal states a desired change in the world plus evidence that would show completion. Vague wishes become engineerable only after success and stopping conditions are observable.

A Goal Is a Testable Destination

A tool is a bounded capability such as search, calculation, or sending a message. The agent decides when and how to request it; the tool returns evidence or an effect.

A Tool Changes What an Agent Can Do

Failures are expected outcomes such as missing data, denied permission, or an unavailable service. A robust graph names them and routes to recovery, escalation, or a safe stop.

Failure Is a Designed State

Feedback is information about the gap between a result and its goal. An agent can use feedback to correct a step, choose another route, or ask a person for help.

Feedback Compares Result With Aim

Nodes, directed edges, state, transitions, and the vocabulary needed to read an agent graph.

Graph and State Literacy

A graph represents things as nodes and relationships as edges. It helps people see possible routes, dependencies, clusters, and loops that a plain list may hide.

Graphs Show Things and Relationships

A system receives information, changes or interprets it, and produces a result. Naming these three parts exposes what a graph node consumes and what downstream nodes may rely on.

Inputs, Transformations, and Outputs

Instructions Constrain Behavior

A learning objective says what a learner should be able to explain, decide, or do. It guides node selection and allows the tutor to check mastery rather than count exposure.

Learning Objectives Describe Demonstrable Change

Memory preserves selected information across steps or sessions. Engineering memory means choosing what to retain, for how long, and who may retrieve it.

Memory Carries Information Forward

Some actions must happen before others because later actions consume earlier results or permissions. A sequence represents that dependency rather than mere visual placement.

Order Can Change an Outcome

Human control means a person can understand, approve, redirect, or stop consequential automation. Control must be available at the moment it can change the outcome.

People Retain Meaningful Control

Private data should be collected and used only for a stated purpose, with minimum necessary detail and a retention rule. Convenience alone does not justify wider access.

Purpose Limits Data Use

An agent may not know whether a claim is true or an action will work. Representing uncertainty enables verification, cautious routing, and useful questions instead of confident guessing.

Learn Before

Related

Learn After